Abstract: Fine-grain content retrieval remains a quasi-unsolved problem in the general case. There are many technical challenges in learning optimal image-to-product models for retrieval systems that efficiently scale to catalogs containing tens of millions of items. This talk details the steps in building such a search platform and highlights the key insights and hurdles found in the process. At Salesforce’s Commerce Cloud Einstein team we have developed a custom multi-stage pipeline of deep metric learning models for product detection, recognition, and reccomendation. We have trained our networks to discover manifolds representing the space of all consumer products with novel techniques. In this talk, we first present state-of-the-art embedding networks and then describe the mechanics of learning an image to product feature map, while also remarking on the most promising new research directions. Implementation level details will be covered insofar as they make a system level understanding of such retrieval possible, and performance evaluation (both statistical as well as qualitative) and metrics used will be described.
Bio: Born and raised in New York, with a Ph.D. in Mathematics and Computer Science (M.S., Ph.D.), Michael has led numerous research and development teams in the ML space. Currently, Michael is a Lead Data Scientist at Salesforce’s Commerce Cloud where he designs new ML technologies powering Salesforce’s recommendation capabilities in e-commerce. From experimental deep neural architectures to ever more performant retrieval methods, Michael enjoys building next-generation predictive analytics technologies. Prior to Salesforce, Mike was a principal engineer on Shutterstock’s ML team, researching models and systems that enable users to more efficiently search and discover content across billions of stock photographs.